The frame in which a human marks out the boundaries of an object makes a huge difference in how well AI software can identify that object through the rest of the video.